Robust Approaches to the Development and Evaluation of Prognostic Classifiers

描述（由申请人提供）：可靠和精确的预后是成功的疾病管理和治疗选择的基础。对于早期发病风险高的患者，可以给予更积极的干预，而对于不太可能对一种治疗产生反应的患者，应考虑其他选择。随着技术的快速进步，广泛的生物和基因组标记物已成为改善疾病预测和治疗结果的潜在工具，并可能导致个性化，量身定制的药物。DNA测序和微阵列等新技术正在生成维度和复杂性呈指数级增长的详细数据。这些数据为准确预测临床结果提供了前所未有的机遇和巨大挑战。为了充分利用这些数据，该提案旨在开发统计方法，以有效地构建和评估疾病风险评估和治疗选择的预后工具。具体来说，在目标1中，我们将通过内核机器回归框架整合复杂的交互效应来开发准确的风险预测模型。我们还将提供非参数程序，用于评估所产生的模型的预测性能。在目标2中，我们提出了使用两阶段研究的新标记物的绝对风险和预测性能的推理程序。在目标3中，我们开发了系统的程序，用于使用患者水平的基线标志物信息来识别可能或可能不会从新治疗中受益的患者亚组。在目标4中，我们专注于高维回归，并开发正则化回归方法来构建回归系数的置信区间和假设检验程序以及估计模型的预测性能。为了增加我们的研究的实际影响，除了创建供公众使用的软件外，我们将应用所提出的程序来预测发生以下疾病的个体风险：（i）使用护士健康研究（NHS）的女性类风湿性关节炎;（ii）使用NHS和健康专业人员随访研究的糖尿病患者心血管疾病;（iii）使用大型免疫遗传学研究的艾滋病病毒感染患者中的艾滋病定义事件;和（iv）CHD或中风，使用妇女健康倡议（WHI）研究。我们还计划开发算法，利用波士顿两家大医院的电子病历（EMR）数据来识别各种自身免疫性疾病的病例。确定的病例将用于相应疾病的后续遗传病例对照研究。这些算法将使EMR临床数据直接用于发现研究。此外，我们将使用随机ACTG临床试验为HIV感染患者制定治疗选择策略，并使用WHI临床试验为预防CVD的饮食干预制定治疗选择策略。将遗传特征、可改变的风险因素与生物标志物沿着纳入风险模型可能会改善临床结果的预测，并最终导致个性化医疗。 公共卫生相关性：该研究提案解决了对先进统计工具的迫切需求，这些工具可以满足当前开发疾病风险和治疗益处预测模型的挑战。通过提供统计工具，使临床研究人员能够有效地制定个性化的疾病管理策略，该提案将加入先前和正在进行的研究活动，以实现寻找高效和具有成本效益的个性化药物的目标。

DESCRIPTION (provided by applicant): A reliable and precise prognosis is fundamental for successful disease management and treatment selection. More aggressive intervention can be given to patients who are at high risk of early disease onset, while patients who are unlikely to respond to one treatment should be considered for alternative options. With the rapid advancement of technology, a wide range of biological and genomic markers have emerged as potential tools for improving the prediction of disease and treatment outcomes, and may lead to personalized, tailored medicine. New technologies such as DNA sequencing and microarrays are generating detailed data with exponentially increasing dimensionality and complexity. These data presents unprecedented opportunities and great challenges for making accurate prediction of clinical outcomes. To take full advantage of such data, this proposal aims to develop statistical approaches to efficiently construct and evaluate prognostic tools for disease risk assessment and treatment selection. Specifically, in Aim 1, we will develop accurate risk prediction models by incorporating complex interactive effects via a kernel machine regression framework. We will also provide non-parametric procedures for assessing the predictive performance of the resulting models. In Aim 2, we propose inference procedures for absolute risks and prediction performance of new markers using two-phase studies. In Aim 3, we develop systematic procedures for identifying subgroups of patients who may or may not benefit from a new treatment using patient level baseline marker information. In Aim 4, we focus on high dimensional regression and develop regularized resampling methods to construct confidence intervals and hypothesis testing procedures for regression coefficients and the prediction performance of estimated models. To increase the practical impact of our research, in addition to creating software for public use, we will apply the proposed procedures to predict individual risk of developing (i) rheumatoid arthritis among women using the Nurse's Health Study (NHS); (ii) CVD among diabetic patients using the NHS and the Health Professional Follow-up Study; (iii) AIDS defining events among HIV infected patients using a large immunogenetic study; and (iv) CHD or stroke using the Women's Health Initiative (WHI) study. We also plan to develop algorithms to identify cases of various autoimmune diseases using electronic medical record (EMR) data from two large hospitals in Boston. The identified cases will be used for subsequent genetic case-control studies of the corresponding diseases. Such algorithms will enable the use of EMR clinical data directly for discovery research. In addition, we will develop treatment selection strategies for HIV infected patients using randomized ACTG clinical trials and for dietary intervention in preventing CVD using WHI clinical trials. Incorporating genetic profile, modifiable risk factors, along with biologic markers into risk models is likely to improve the prediction of clinical outcomes and ultimately lead to personalized medicine. PUBLIC HEALTH RELEVANCE: The research proposal addresses the pressing need for advanced statistical tools that meet challenges in current development of prediction models for disease risk and treatment benefit. By providing statistical tools that enable clinical investigators to effectively develop personalized disease management strategies, this proposal will join prior and ongoing research activities towards the goal of finding efficient and cost effective personalized medicine.